Anthropic: Claude Code v2.1.178 Introduces Tool Parameter Matching in Permissions and Nested Skills

Anthropic has released Claude Code v2.1.178, a CLI release for agentic programming that brings finer-grained permissions and better skills management.

How do the new per-parameter tool permissions work?

The new feature is the Tool(param:value) syntax for permission rules, with wildcard support. This lets permissions be tied to specific call parameters rather than just to the tool name. For example, the rule Agent(model:opus) can block subagents from being launched on the Opus model. Such granularity lets teams exercise precise control over which actions an agent is allowed to take.

What changes with skills and subagents?

Skills placed in nested .claude/skills/ directories now load automatically, with a prefix of the form <dir>:<name> that resolves name collisions. In auto mode, launching subagents additionally passes through a classifier before the launch, filtering out unnecessary or risky calls. The /doctor command has gained a cleaner “flat tree” view of status.

What are the key bug fixes?

The release removes several annoying bugs: an OOM (out-of-memory) crash when the CLI inherits a stale websocket or OAuth environment variable, an issue with Claude connecting in Chrome with another account’s OAuth token, and background sessions that could remain stuck at “Working” status indefinitely. Also fixed: the Esc key in VSCode no longer cancels the CJK IME character selection window.